McOmet: Multimodal Fusion Transformer for Physical Audiovisual Commonsense Reasoning

Physical commonsense reasoning is essential for building reliable and interpretable AI systems, which involves a general understanding of the physical properties and affordances of everyday objects, how these objects can be manipulated, and how they interact with others. It is fundamentally a multi-modal task, as physical properties are manifested through multiple modalities, including vision and acoustics. In this work, we present a unified framework, named Multimodal Commonsense Transformer (MCOMET), for physical audiovisual commonsense reasoning. MCOMET has two intriguing properties: i) it fully mines higher-ordered temporal relationships across modalities (e.g., pairs, triplets, and quadruplets); and ii) it restricts the cross-modal flow through the feature collection and propagation mechanism along with tight fusion bottlenecks, forcing the model to attend the most relevant parts in each modality and suppressing the dissemination of noisy information. We evaluate our model on a very recent public benchmark, PACS. Results show that MCOMET significantly outperforms a variety of strong baselines, revealing powerful multi-modal commonsense reasoning capabilities.